K-Nearest Neighbors (KNN) Regression as a Tool for Failure Rate Prediction of Si MOSFET

Abstract

Forecasting the operational lifetime of a Metal-Oxide-Semiconductor Field-Effect Transistor (MOSFET) is crucial for ensuring the stability and robustness of electronic systems. These devices experience temperature cycling, voltage stressing, and high-frequency switching over time, which subsequently degrades their electrical characteristics, including threshold voltage, on-resistance, and gate charge. If these changes go unchecked, they may result in issues that compromise the control and safety of the entire system. Life prediction enhances the performance of electronic systems by focusing on the mitigation level of their failing subsystems, thereby improving overall efficiency. The lifetime of a MOSFET can be determined by tracking the drain-to-source ON resistance (RDS(on)) curve over its lifespan. The experimental result of the proposed system at a power level of 1100W with a regulated output voltage of 211 VDC has an output voltage ripple of ~ 4.256 %, and the efficiency of the system is 93.51 %. The K-Nearest Neighbors (KNN) Regression method serves to estimate the RDS(on) variability and predict well in advance. It utilizes a deep learning model that is trained on a provided dataset encompassing the lifecycle of power MOSFETs. The results obtained are highly optimistic, indicating that the proposed method is efficient. The presented method achieves over 99 % training efficiency. When evaluating this predictive model, the root mean squared error (RMSE) was at 0.0006, alongside a 0.9987 R2 score.